EIF2AK1 As a Key Genetic Determinant of Luspatercept Resistance in Transfusion‐Dependent Thalassemia

To the Editor,

1.

Transfusion‐dependent beta‐thalassemia (TDT) is characterized by ineffective erythropoiesis, chronic hemolytic anemia, and lifelong dependence on red blood cell (RBC) transfusions [1, 2]. Despite advances in supportive care, TDT remains associated with substantial morbidity and mortality, primarily due to iron overload and related complications, which require complex and lifelong management strategies [1, 3, 4].

Luspatercept, a recombinant fusion protein that acts as a ligand‐trap for selected members of the transforming growth factor‐β (TGF‐β) superfamily, enhances late‐stage erythroid maturation and was the first approved drug to reduce the transfusion burden in patients with TDT [5, 6]. Clinical trials have demonstrated that luspatercept can significantly reduce transfusion requirements in a proportion of patients. Nevertheless, only approximately 40%–50% of patients respond to treatment in real‐world settings, leaving a substantial proportion of individuals without clinical benefit [6, 7, 8], and the mechanisms underlying this interindividual variability are poorly understood. Given the biological complexity of erythropoiesis and the interplay between genetic and regulatory factors, genetic variability may contribute to differences in drug response among patients.

We herein conducted an exome‐wide association study in 70 TDT patients, stratified by drug response, to identify rare and low‐frequency variants that may influence either erythroid maturation or the molecular pathways involved in the response to luspatercept. Detailed methods are provided in the Supporting Information.

The study cohort included 51 TDT patients (discovery set) and an independent cohort of 19 TDT patients (validation set) enrolled at two hospitals (Table S1). Molecular diagnosis was established by sequencing HBB, HBA1, and HBA2 using the Devyser Thalassemia NGS kit (Devyser, Sweden) (Table S2). All patients received luspatercept, and treatment response was assessed across any 12‐week interval. Baseline laboratory parameters were collected before luspatercept initiation, whereas responder status was assigned according to treatment outcome. Patients were classified into age‐ and sex‐matched responders (R; ≥ 33% reduction in transfusion burden during any 12‐week period) or non‐responders (NR) [6]. While response classification was harmonized across centers, differences in pre‐luspatercept management limited cross‐cohort clinical comparability (Table S3); therefore, the validation set was primarily used to replicate the genetic association. Based on the ≥ 33% reduction in transfusion burden cutoff, 19/51 (37.3%) patients in the discovery set were classified as R, while 32 (62.7%) were classified as NR, consistent with previously published data [6]. A similar proportion of R/NR was observed in the validation set, with 5/19 (26.3%) patients classified as R and 14/19 (73.7%) as NR (p = 0.57).

Building on this framework, we performed an exome‐wide association analysis in the discovery cohort to identify genetic determinants of treatment outcome. Genomic DNA extraction from peripheral blood, whole‐exome sequencing, and variant calling were performed as previously described [9, 10]. A total of 123 836 unique variants were identified in the discovery set. After linkage disequilibrium pruning (r ² > 0.8) and gene‐level burden analysis, 635 genes showed significant associations with treatment response, with 354 linked to R and 281 linked to NR (Figure 1A). To gain insight into the biological processes underlying these associations, we performed functional enrichment analyses on the genes associated with response status. Gene Ontology enrichment analysis revealed an overrepresentation of the biological process “response to iron ion starvation” (GO:0010109) among genes associated with NR (Figure 1B). Consistently, Reactome pathway analysis highlighted enrichment of the heme‐deficiency‐induced EIF2AK1 signaling pathway (p < 0.05) (Figure 1B). Taken together, these results revealed dysregulated erythroid stress signaling as a potential biological background associated with reduced responsiveness to luspatercept.

FIGURE 1.

Analysis of gene mutational load and functional enrichment. (A) Gene‐level burden test performed by considering the number of variants observed per gene. The volcano plot displayed genes as a function of the model‐estimated effect size (β, x‐axis) and their statistical significance expressed as −log10(p‐value) (y‐axis). The β coefficient represented the direction and the association magnitude between mutational load and clinical response: Genes significantly enriched in responders (R; β > 0 and p < 0.05) were shown in red, those significantly enriched in non‐responders (NR; β < 0 and p < 0.05) were shown in blue, while not significant genes were shown in gray. (B) Functional enrichment analysis performed on NR‐enriched genes. The lollipop plot showed the significantly enriched Reactome pathway (R‐HSA, in light blue) and Gene Ontology (GO, in dark blue) found by the burden test. The length of the bars represented fold enrichment, while the size of the lollipop head indicated the number of genes involved in the GO and pathway, respectively. Statistical significance was indicated by increasing color intensity. Pathways of interest were labeled. (C) Violin plots showing the effect of the eQTL rs12657 on EIF2AK1 expression in EBV‐transformed lymphocytes (p < 0.0001; normalized effect size [NES] = 0.74) and in spleen tissue (p = 0.26; NES = 0.11). The median normalized expression value for each genotypic class is indicated (data retrieved from the Genotype‐Tissue Expression (GTEx) database, https://www.gtexportal.org/home/). (D) Violin plot showing EIF2AK1 expression in reticulocytes from non‐responder patients, stratified by EIF2AK1 genotype, non‐carriers (GG) and carriers (GA). Variant carriers showed significantly increased normalized expression compared with non‐carriers (p = 0.0010, Mann–Whitney test).

To narrow the analysis to variants potentially relevant to non‐response, we next performed variant‐level prioritization. Overall, 189 variants were identified as enriched in R (p < 0.01, log₂OR > 0), whereas 123 variants were identified as enriched in NR (p < 0.01, log₂OR < 0). To identify biologically and potentially clinically relevant variants, we prioritized rare and low‐frequency variants in the general population (alternative allele frequency [AF] < 0.05), selecting those absent in R. This analysis highlighted a recurrent EIF2AK1 3′UTR variant (NM_014413.4:c.*630C>T; rs12657; AF = 0.04670), observed in 31% of NR patients (10/32) and absent among R subjects (p = 0.008). This was validated in an independent cohort by PCR amplification and direct sequencing. Primer sequences are available upon request. In the validation set, we confirmed the pattern observed in the discovery set, with the variant occurring exclusively among NR; however, the association did not reach statistical significance (p = 0.12), likely due to the limited sample size. When cohorts were combined, variant carriage was enriched among NR (13/46, 28.3%) and was associated with a markedly increased likelihood of non‐response to luspatercept (OR 20.6; 95% CI 2.5–2.7 × 10 [3]; p = 0.0016). Neither the main effect of cohort nor the variant‐by‐cohort interaction was statistically significant (p = 0.32 and p = 0.45), indicating no evidence that the variant effect differed between datasets. In the discovery cohort, variant carriage was strongly associated with non‐response (OR ≈ 18), with all carriers observed among NR. In the validation cohort, the effect was directionally consistent (OR ≈ 3), although estimated with lower precision because of the smaller sample size. The meta‐analysis yielded a combined OR ≈ 8 and showed no evidence of between‐cohort heterogeneity, supporting the discovery finding and indicating that the association is consistent across datasets.

Comparison with European‐ancestry gnomAD data revealed a significant trend of increasing A‐allele frequency from population controls to β‐thalassemia patients and further to the NR subgroup (Z = 3.47, p = 0.00052, Armitage trend test) (Table S4). This pattern suggests that rs12657 is not simply enriched in the disease background but may instead identify a subset of patients characterized by reduced responsiveness to treatment.

To explore the functional relevance of rs12657, we first interrogated public expression data. In silico analysis identified rs12657 as an expression quantitative trait locus (eQTL), with the A allele linked to increased EIF2AK1 expression (Figure 1C). Moreover, we assessed EIF2AK1 mRNA levels in β‐thalassemia patients stratified by variant genotype. mRNA was isolated from reticulocytes as previously described [11], and EIF2AK1 expression was quantified by digital PCR analysis [12]. Consistently, variant carriers exhibited significantly higher basal levels of EIF2AK1 transcript in reticulocytes compared with individuals homozygous for the reference allele (Figure 1D), supporting a functional impact of the variant on gene expression. Overall, EIF2AK1, which encodes the heme‐regulated inhibitor kinase (HRI), a central regulator of erythroid stress signaling [12], emerged as a candidate biomarker of treatment non‐response.

We next examined the baseline clinical features of the discovery set associated with the EIF2AK1 variant. The association between the variant and treatment outcome was evaluated using Firth's penalized logistic regression to account for the absence of variant carriers among R and the small sample size. Ten NR patients carrying the variant (NR‐var) were compared with NR non‐carriers and R. To account for differences in transfusion burden, iron intake was used as a quantitative proxy of transfusion exposure and included as an additional covariate in the comparative analyses. Several laboratory parameters showed an ordered trend across NR‐var, NR, and R groups, suggesting a distinct NR‐var profile (Table 1). Specifically, before treatment initiation, NR‐var patients exhibited higher erythrocyte hemoglobin content, reflected by increased mean corpuscular hemoglobin, together with higher circulating iron indices, including serum iron and transferrin saturation (TSAT), which progressively decreased across NR and R groups. After adjustment for iron intake, differences in selected erythroid indices and iron metabolism biomarkers remained detectable, supporting a phenotype not entirely explained by transfusion burden (Table 1). Additionally, HbF levels declined from R to NR‐var, and a similar trend was observed for HbA2. However, these differences were no longer significant after accounting for variability in iron intake. Taken together, these parameters revealed a distinct baseline erythroid and iron‐related profile in NR‐var patients.

TABLE 1.

Comparison of baseline hematological and biochemical parameters among NR‐var, NR, and R groups.

Variable	Unit	NR‐var			NR			R			P a	P1 a	P2 a	P3 a	P_adj. iron intake b
		n	Median	IQR	n	Median	IQR	n	Median	IQR
Age	years	10	47.00	40.2–51.8	22	43.00	32.2–53.5	19	36.00	30.5–48.0	0.2223	—	—	—	—
RBC	106/μL	10	3.18	3.0–3.3	22	3.42	3.1–3.6	19	3.62	3.3–3.8	0.0040	—	**	—	0.0320
Hemoglobin	g/dL	10	9.20	8.9–9.6	22	10.10	8.9–10.4	19	9.50	8.6–10.2	0.2990	—	—	—	0.4571
Hematocrit	%	10	26.10	25.4–27.2	22	28.40	25.6–29.7	19	28.10	25.3–29.1	0.3097	—	—	—	0.5135
MCV	fL	10	83.15	82.5–84.0	22	82.10	80.5–83.4	19	79.60	74.7–83.7	0.0557	—	—	—	0.0681
MCH	pg	10	29.10	28.9–29.4	22	29.05	28.6–29.5	19	27.30	25.1–28.8	0.0043	—	*	*	0.0090
MCHC	g/dL	10	35.20	34.8–35.4	22	35.45	34.7–35.8	19	34.00	33.8–35.0	0.0127	—	—	**	0.1374
RDW	%	10	15.80	15.1–17.1	22	15.45	14.5–17.5	18	18.30	15.5–24.1	0.0123	—	—	*	0.0348
Reticulocytes	%	9	0.96	0.8–1.7	22	0.72	0.4–1.9	18	2.48	0.5–6.7	0.1013	—	—	—	0.1350
ARC	103/μL	10	35.00	26.7–81.0	22	22.5	11.2–67.3	17	75.0	18.0–234.0	0.1097	—	—	—	0.8869
WBC	103/μL	10	8.30	6.6–11.2	22	7.63	6.5–10.8	19	7.90	5.5–10.1	0.9186	—	—	—	0.5263
PLT	103/μL	10	523.00	321.8–671.2	22	295.00	191.0–402.2	19	298.00	152.0–488.0	0.0320	—	*	—	0.0015
MPV	fL	9	10.60	10.2–10.7	16	10.70	10.1–11.8	11	10.50	10.1–11.2	0.6774	—	—	—	0.6196
Total Bilirubin	mg/dL	9	2.40	1.8–2.7	22	2.28	1.4–2.7	17	1.97	1.7–3.2	0.6671	—	—	—	0.3067
Unc bilirubin	mg/dL	8	1.70	1.1–1.9	21	1.43	0.8–1.8	16	1.58	1.1–2.7	0.2457	—	—	—	0.1502
LDH	U/L	7	232.00	175.5–241.5	22	167.00	153.0–218.0	19	195.00	168.5–289.0	0.2022	—	—	—	0.4934
Haptoglobin	mg/dL	9	43.00	13.0–60.0	20	30.50	11.5–41.0	14	20.50	0.0–34.0	0.1856	—	—	—	0.4074
AST	U/L	8	28.50	28.0–31.0	21	22.00	19.0–28.0	16	25.00	21.0–29.2	0.1821	—	—	—	0.7550
ALT	U/L	10	20.00	14.2–31.8	22	21.00	15.2–27.8	19	16.00	13.5–21.0	0.4653	—	—	—	0.9521
Serum Iron	μg/dL	10	305.00	291.8–320.5	22	260.50	212.8–316.0	19	208.00	171.5–228.5	0.0005	—	***	*	0.0001
Ferritin	ng/mL	10	553.00	483.5–871.8	22	680.00	390.0–930.8	19	932.00	509.5–2036.5	0.4003	—	—	—	0.3400
Transferrin	mg/dL	10	199.50	180.8–206.2	22	172.50	159.5–200.0	19	166.00	146.5–191.0	0.1063	—	—	—	0.0523
TSAT	%	10	110.50	100.2–119.5	22	99.00	91.5–112.2	19	87.00	72.5–98.0	0.0020	—	**	*	0.0019
LIC	mg/gdw	10	2.42	1.7–4.1	22	1.69	1.4–4.1	19	3.99	1.7–9.9	0.0599	—	—	—	0.0151
HbA2	%	10	2.40	2.3–2.5	22	2.50	2.4–2.7	19	2.70	2.5–3.2	0.0407	—	*	—	0.2690
HbF	%	10	3.15	2.5–6.6	22	3.55	2.9–6.0	19	10.40	7.8–15.8	0.0010	—	**	**	0.0971

Note: Bold p‐values indicate statistical significance p < 0.05.

Abbreviations: ALT, alanine aminotransferase; ARC, absolute reticulocyte count; AST, aspartate aminotransferase; IQR, interquartile range; LDH, lactate dehydrogenase; LIC, liver iron concentration; MCH, mean corpuscular hemoglobin; MCHC, mean corpuscular hemoglobin concentration; MCV, mean corpuscular volume; n, number of observations; NR, non‐responders without EIF2AK1 variant; NR‐var, non‐responders carrying EIF2AK1 variant; R, responders; RBC, red blood cells; RDW, red cell distribution width; TSAT, transferrin saturation.

^a p‐value by Kruskal–Wallis test and Dunn's post hoc analysis for multiple comparisons (P1, NR‐var versus NR; P2 NR‐var versus R; P3, NR versus R).

^b Nominal iron‐intake‐adjusted p‐value from linear regression models including clinical group and pre‐treatment iron intake as covariates.

^* p < 0.05.

^** p < 0.01.

^*** p < 0.001.

To further assess whether this phenotype extended beyond the response‐based comparison, we stratified the expanded cohort by EIF2AK1 genotype, irrespective of treatment response. At baseline, heterozygous subjects exhibited higher serum iron and TSAT and lower RBC, HbF, and HbA2 levels compared with non‐carriers (Figure 2A). When assessed in multivariable linear regression models adjusted for age, sex, splenectomy status, iron chelation, and iron intake, serum iron emerged as the parameter most strongly associated with the EIF2AK1 variant (Figure 2B and Table S5). Overall, these findings suggested that EIF2AK1 may represent not only a candidate marker of treatment non‐response, but also an indicator of a distinct clinical and biochemical β‐thalassemia phenotype.

FIGURE 2.

Association between EIF2AK1 carrier status and clinical parameters. (A) Representative boxplots of selected clinical parameters stratified by EIF2AK1 genotype (carriers—GA, blue; non‐carriers—GG, red). Boxes represent the median and interquartile range (IQR), with whiskers extending to 1.5 × IQR. p‐values by Mann–Whitney test. (B) Lollipop plot summarizing the results of multivariable linear regression models fitted for each continuous clinical outcome, with carrier status as the primary predictor and adjustment for age, sex, splenectomy status, iron intake, and iron chelation. The x‐axis reports the nominal p‐value for the EIF2AK1 term (x‐axis reversed so smaller p‐values appear to the right); the dashed vertical line indicates the p = 0.05 threshold. Dot color intensity reflects the Benjamini‐Hochberg false discovery rate (FDR) across outcomes (darker dots indicate smaller FDR).

In β‐thalassemia, where ineffective erythropoiesis, oxidative stress, globin imbalance, and iron dysregulation are tightly interconnected, variation in the HRI pathway may influence the ability of erythroid precursors to respond to agents that promote terminal maturation. This interpretation is supported by evidence linking HRI activity to β‐thalassemia pathophysiology [13, 14], regulation of HbF through the eIF2α‐ATF4‐BCL11A axis [15], and modulation of disease severity under stress conditions [16]. In addition, several genome‐wide association studies have linked common variants within the EIF2AK1 locus to erythroid hematological traits (Table S6). Our results support a model in which increased basal HRI activity defines a stress‐adapted erythroid state that may be less permissive to efficient terminal maturation under luspatercept treatment. Within this framework, the EIF2AK1 3′UTR variant may act as a marker of this biological state, identifying a subset of patients with reduced responsiveness to erythroid maturation therapy. Mechanistically, elevated basal HRI expression may sustain chronic eIF2α‐ATF4 signaling, promoting maladaptive dyserythropoiesis and maintaining erythroblasts in a stress‐adapted, poorly hemoglobinizing state, thereby limiting terminal differentiation under treatment. Although this model remains speculative, it is consistent with the known biology of the HRI pathway and provides a coherent explanation for the observed non‐responder phenotype. Despite the study limitations, including the limited number of variant carriers and the lack of functional validation of HRI overactivation, our findings may have relevant implications for patient stratification and treatment optimization in TDT. If confirmed in larger independent cohorts, EIF2AK1 genotyping may help identify patients less likely to benefit from luspatercept and support more individualized therapeutic decision‐making. Given the link between HRI signaling and iron homeostasis, EIF2AK1 may also deserve further investigation as a potential modifier of iron chelation requirements or response.

Author Contributions

All authors read and approved the final version of the article. R.R., I.M., and I.A. designed the study. A.I., F.R., and B.E.R. performed the statistical analysis. V.B. and I.V. performed gDNA extraction, mRNA isolation, and gene expression analysis. R.M. and A.N. performed molecular genetics analysis, library enrichment, and NGS analysis. V.A.L. and M.C. performed bioinformatic analysis for NGS data. S.L., N.S., D.L.P., T.C., C.M.G., N.K., I.M., and G.B.F. cared for the patients. A.I. and F.R. wrote the manuscript. R.R., I.M., I.A., M.C., and G.B.F. carried out a critical revision of the manuscript.

Funding

This research was funded by the Italian Ministry of University and Research PRIN 2022 to R.R. and I.M. (2022L5YWZT). This work is generated within the European Reference Network on Rare Hematological Diseases (ERN‐EuroBloodNet). R.R., I.M., and I.A. are members of EuroBloodNet.

Ethics Statement

The local University Ethical Committee approved the collection of patient data (Comitato Etico Campania 3 – 283/2024).

Consent

DNA samples were obtained from patients after they had signed their informed consent, in accordance with the Declaration of Helsinki.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Table S1: Baseline clinical features of non‐responder (NR) and responder (R) patients in the discovery and validation sets.

Table S2: Globin gene genotypes in the enrolled patients.

Table S3: Baseline clinical characteristics of non‐responder (NR) and responder (R) patients in the discovery and validation sets.

Table S4: Allele and genotype frequencies of EIF2AK1 rs12657 variant.

Table S5: Genotype–phenotype associations by EIF2AK1 carrier status.

Table S6: Genetic associations between EIF2AK1 (7p22.1) variants and erythroid hematological traits.

Data Availability Statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.